Privacy-preserving AI refers to methods that limit how much personal or sensitive information is revealed while AI systems are trained, operated, or evaluated. Depending on the use case, techniques can include data minimisation, removing or masking identifiers, processing data locally, secure computation, and differential privacy, which adds carefully calibrated noise to limit what can be inferred about individuals.
These safeguards matter because AI workflows may expose information through training data, prompts, model outputs, or logs. In real deployments, teams can reduce risk by collecting only necessary data, restricting access, testing for memorisation or leakage, and setting retention controls. No single technique guarantees privacy: protections need to be assessed against the data, threat model, and system design, and balanced with utility and security requirements.